[PATCH] iommu/arm-smmu-v3: Allow stream table to have nodes with the same ID

[Pranjal Shrivastava]

On Fri, Apr 11, 2025 at 10:01:55AM -0300, Jason Gunthorpe wrote:
> On Fri, Apr 11, 2025 at 01:10:29PM +0100, Robin Murphy wrote:
> 
> > This is adding support for StreamID aliasing between devices, and as such it
> > is incomplete. It's not OK to just allow devices to arbitrarily rewrite each
> > other's STEs,
> 
> Okay, yes, we should be checking the iommu_group before permitting two
> devices to share the STE. That is an easy fix, see below
> 
> > Aliases can only be permitted within a group, which means
> > arm_smmu_device_group() also has to check and account for them in the first
> > place - note that that applies to PCI devices as well, because as
> > soon as we
> 
> On this system the alias come from the PCI DMA alias support and
> pci_device_group() is already correctly grouping things.
> 
> Aliases from other places, like the IORT, never did work..
> 
> > allow StreamID aliasing at all then we're inherently allowing RID->SID
> > mappings to alias outside the PCI hierarchy in ways that pci_device_group()
> > can't know about. It should work out basically the same as SMMUv2, just with
> > the streams tree in place of the S2CR array.
> 
> You mean the logic in v2's arm_smmu_device_group() to consult the
> stream map to select the group if the IORT is creating aliases? Yes it
> could be done..
> 
> However, this is a significant regression fix and I think we can be
> confident there are no IORT tables in the wild that have aliases or
> they would already be broken.
> 
> How about we add the missing validation that the group is the same,
> that is easy to do and should be there anyhow:
> 
> static int arm_smmu_streams_cmp_node(struct rb_node *lhs,
> 				     const struct rb_node *rhs)
> {
> 	struct arm_smmu_stream *stream_lhs =
> 		&rb_entry(lhs, struct arm_smmu_stream, node);
> 	struct arm_smmu_stream *stream_rhs =
> 		rb_entry(rhs, struct arm_smmu_stream, node);
> 
> 	if (stream_lhs->id < stream_rhs->id)
> 		return -1;
> 	if (stream_lhs->id > stream_rhs->id)
> 		return 1;
> 
> 	/*
> 	 * The stream table can have multiple nodes with the same ID if there
> 	 * are DMA aliases. If multiple masters share the same iommu group then
> 	 * they can use the overlapping STEs within the group.
> 	 */
> 	if (stream_lhs->master->dev->iommu_group ==
> 	    stream_rhs->master->dev->iommu_group) {
> 		if (stream_lhs < stream_rhs)
> 			return -1;
> 		if (stream_lhs > stream_rhs)
> 			return 1;
> 	}
> 	return 0;
> }
> 
> That change will narrow this patch to only enable PCI DMA aliases that
> already generate the correct iommu groupings. Other sources of alising
> that don't generate the right groupings will continue to fail as they
> do today.
> 

Isn't the device grouped *after* the ops->probe_device call? I see that
the dev->iommu_group is assigned the ops->probe_device call in
iommu_init_device.. so I guess this would still fail?

Thanks,
Praan

> Then I propose continuing to wait for a user before adding support for
> more alias scenarios to arm_smmu_device_group()?
> 
> > > +	/*
> > > +	 * If there are DMA alises then there are multiple devices with the same
> > > +	 * stream ID and we cannot reliably convert from SID to master.
> > > +	 */
> > > +	if (node->rb_left &&
> > > +	    rb_entry(node->rb_left, struct arm_smmu_stream, node)->id == sid)
> > > +		return NULL;
> > > +	if (node->rb_right &&
> > > +	    rb_entry(node->rb_right, struct arm_smmu_stream, node)->id == sid)
> > > +		return NULL;
> > 
> > This doesn't really work - the whole mechanism needs to fundamentally change
> > to mapping StreamIDs to groups rather than to devices. Then it's really up
> > to individual callers what they want to do if the group has more than one
> > device.
> 
> There are only two callers. One is using it to print the log message,
> in this case it will fall back to the unknown stream ID path and still
> print a log message. This could perhaps print the group ID # instead
> of the raw stream ID but I wouldn't do that in a regression fix rc
> patch.
> 
> The other is doing stall/future PRI, and I don't think we should be
> doing iommu_group based fault reporting at all. Returning NULL
> effectively disables it.
> 
> Jason